A novel pooling layer based on gaussian function with wavelet transform

Abstract

Convolution represent basic layer in the convolutional neural network, but it can result in big size of the data, which may increase the complexity of the network. Different pooling methods are used to perform down sample these data. In this paper, we have proposed a novel pooling method by using Gaussian function to determine the wavelet filter coefficients. At first, the basic statistics are determined for each pool size of the signal, then Gaussian probability distribution function is determined. According to the procedure of extracting the features, three methods are proposed, the first method is used the normalized values of basic statistics as wavelet filter to be multiplied by original signal, the second method used the determined statistics as features of the original signal, then multiplied it with constant wavelet filter based on Gaussian, while the third method is similar to first method, except it depend on entire signal instead of each pool size. The proposed methods are combined with other standard methods such as max and pooling. The experiments are performed on different datasets and the results show that the proposed methods perform or outperform other methods and can increase performance of the (CNN).